1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
2
3 /* Authors: Bernard Metzler <bmt@zurich.ibm.com> */
4 /* Copyright (c) 2008-2019, IBM Corporation */
5
6 #include <linux/errno.h>
7 #include <linux/types.h>
8 #include <linux/net.h>
9 #include <linux/scatterlist.h>
10 #include <linux/highmem.h>
11 #include <net/tcp.h>
12
13 #include <rdma/iw_cm.h>
14 #include <rdma/ib_verbs.h>
15 #include <rdma/ib_user_verbs.h>
16
17 #include "siw.h"
18 #include "siw_verbs.h"
19 #include "siw_mem.h"
20
21 #define MAX_HDR_INLINE \
22 (((uint32_t)(sizeof(struct siw_rreq_pkt) - \
23 sizeof(struct iwarp_send))) & 0xF8)
24
siw_get_pblpage(struct siw_mem * mem,u64 addr,int * idx)25 static struct page *siw_get_pblpage(struct siw_mem *mem, u64 addr, int *idx)
26 {
27 struct siw_pbl *pbl = mem->pbl;
28 u64 offset = addr - mem->va;
29 dma_addr_t paddr = siw_pbl_get_buffer(pbl, offset, NULL, idx);
30
31 if (paddr)
32 return ib_virt_dma_to_page(paddr);
33
34 return NULL;
35 }
36
siw_get_page(struct siw_mem * mem,struct siw_sge * sge,unsigned long offset,int * pbl_idx)37 static struct page *siw_get_page(struct siw_mem *mem, struct siw_sge *sge,
38 unsigned long offset, int *pbl_idx)
39 {
40 if (!mem->is_pbl)
41 return siw_get_upage(mem->umem, sge->laddr + offset);
42 else
43 return siw_get_pblpage(mem, sge->laddr + offset, pbl_idx);
44 }
45
46 /*
47 * Copy short payload at provided destination payload address
48 */
siw_try_1seg(struct siw_iwarp_tx * c_tx,void * paddr)49 static int siw_try_1seg(struct siw_iwarp_tx *c_tx, void *paddr)
50 {
51 struct siw_wqe *wqe = &c_tx->wqe_active;
52 struct siw_sge *sge = &wqe->sqe.sge[0];
53 u32 bytes = sge->length;
54
55 if (bytes > MAX_HDR_INLINE || wqe->sqe.num_sge != 1)
56 return MAX_HDR_INLINE + 1;
57
58 if (!bytes)
59 return 0;
60
61 if (tx_flags(wqe) & SIW_WQE_INLINE) {
62 memcpy(paddr, &wqe->sqe.sge[1], bytes);
63 } else {
64 struct siw_mem *mem = wqe->mem[0];
65
66 if (!mem->mem_obj) {
67 /* Kernel client using kva */
68 memcpy(paddr, ib_virt_dma_to_ptr(sge->laddr), bytes);
69 } else if (c_tx->in_syscall) {
70 if (copy_from_user(paddr, u64_to_user_ptr(sge->laddr),
71 bytes))
72 return -EFAULT;
73 } else {
74 unsigned int off = sge->laddr & ~PAGE_MASK;
75 struct page *p;
76 char *buffer;
77 int pbl_idx = 0;
78
79 p = siw_get_page(mem, sge, 0, &pbl_idx);
80 if (unlikely(!p))
81 return -EFAULT;
82
83 buffer = kmap_local_page(p);
84
85 if (likely(PAGE_SIZE - off >= bytes)) {
86 memcpy(paddr, buffer + off, bytes);
87 } else {
88 unsigned long part = bytes - (PAGE_SIZE - off);
89
90 memcpy(paddr, buffer + off, part);
91 kunmap_local(buffer);
92
93 p = siw_get_page(mem, sge, part, &pbl_idx);
94 if (unlikely(!p))
95 return -EFAULT;
96
97 buffer = kmap_local_page(p);
98 memcpy(paddr + part, buffer, bytes - part);
99 }
100 kunmap_local(buffer);
101 }
102 }
103 return (int)bytes;
104 }
105
106 #define PKT_FRAGMENTED 1
107 #define PKT_COMPLETE 0
108
109 /*
110 * siw_qp_prepare_tx()
111 *
112 * Prepare tx state for sending out one fpdu. Builds complete pkt
113 * if no user data or only immediate data are present.
114 *
115 * returns PKT_COMPLETE if complete pkt built, PKT_FRAGMENTED otherwise.
116 */
siw_qp_prepare_tx(struct siw_iwarp_tx * c_tx)117 static int siw_qp_prepare_tx(struct siw_iwarp_tx *c_tx)
118 {
119 struct siw_wqe *wqe = &c_tx->wqe_active;
120 char *crc = NULL;
121 int data = 0;
122
123 switch (tx_type(wqe)) {
124 case SIW_OP_READ:
125 case SIW_OP_READ_LOCAL_INV:
126 memcpy(&c_tx->pkt.ctrl,
127 &iwarp_pktinfo[RDMAP_RDMA_READ_REQ].ctrl,
128 sizeof(struct iwarp_ctrl));
129
130 c_tx->pkt.rreq.rsvd = 0;
131 c_tx->pkt.rreq.ddp_qn = htonl(RDMAP_UNTAGGED_QN_RDMA_READ);
132 c_tx->pkt.rreq.ddp_msn =
133 htonl(++c_tx->ddp_msn[RDMAP_UNTAGGED_QN_RDMA_READ]);
134 c_tx->pkt.rreq.ddp_mo = 0;
135 c_tx->pkt.rreq.sink_stag = htonl(wqe->sqe.sge[0].lkey);
136 c_tx->pkt.rreq.sink_to =
137 cpu_to_be64(wqe->sqe.sge[0].laddr);
138 c_tx->pkt.rreq.source_stag = htonl(wqe->sqe.rkey);
139 c_tx->pkt.rreq.source_to = cpu_to_be64(wqe->sqe.raddr);
140 c_tx->pkt.rreq.read_size = htonl(wqe->sqe.sge[0].length);
141
142 c_tx->ctrl_len = sizeof(struct iwarp_rdma_rreq);
143 crc = (char *)&c_tx->pkt.rreq_pkt.crc;
144 break;
145
146 case SIW_OP_SEND:
147 if (tx_flags(wqe) & SIW_WQE_SOLICITED)
148 memcpy(&c_tx->pkt.ctrl,
149 &iwarp_pktinfo[RDMAP_SEND_SE].ctrl,
150 sizeof(struct iwarp_ctrl));
151 else
152 memcpy(&c_tx->pkt.ctrl, &iwarp_pktinfo[RDMAP_SEND].ctrl,
153 sizeof(struct iwarp_ctrl));
154
155 c_tx->pkt.send.ddp_qn = RDMAP_UNTAGGED_QN_SEND;
156 c_tx->pkt.send.ddp_msn =
157 htonl(++c_tx->ddp_msn[RDMAP_UNTAGGED_QN_SEND]);
158 c_tx->pkt.send.ddp_mo = 0;
159
160 c_tx->pkt.send_inv.inval_stag = 0;
161
162 c_tx->ctrl_len = sizeof(struct iwarp_send);
163
164 crc = (char *)&c_tx->pkt.send_pkt.crc;
165 data = siw_try_1seg(c_tx, crc);
166 break;
167
168 case SIW_OP_SEND_REMOTE_INV:
169 if (tx_flags(wqe) & SIW_WQE_SOLICITED)
170 memcpy(&c_tx->pkt.ctrl,
171 &iwarp_pktinfo[RDMAP_SEND_SE_INVAL].ctrl,
172 sizeof(struct iwarp_ctrl));
173 else
174 memcpy(&c_tx->pkt.ctrl,
175 &iwarp_pktinfo[RDMAP_SEND_INVAL].ctrl,
176 sizeof(struct iwarp_ctrl));
177
178 c_tx->pkt.send.ddp_qn = RDMAP_UNTAGGED_QN_SEND;
179 c_tx->pkt.send.ddp_msn =
180 htonl(++c_tx->ddp_msn[RDMAP_UNTAGGED_QN_SEND]);
181 c_tx->pkt.send.ddp_mo = 0;
182
183 c_tx->pkt.send_inv.inval_stag = cpu_to_be32(wqe->sqe.rkey);
184
185 c_tx->ctrl_len = sizeof(struct iwarp_send_inv);
186
187 crc = (char *)&c_tx->pkt.send_pkt.crc;
188 data = siw_try_1seg(c_tx, crc);
189 break;
190
191 case SIW_OP_WRITE:
192 memcpy(&c_tx->pkt.ctrl, &iwarp_pktinfo[RDMAP_RDMA_WRITE].ctrl,
193 sizeof(struct iwarp_ctrl));
194
195 c_tx->pkt.rwrite.sink_stag = htonl(wqe->sqe.rkey);
196 c_tx->pkt.rwrite.sink_to = cpu_to_be64(wqe->sqe.raddr);
197 c_tx->ctrl_len = sizeof(struct iwarp_rdma_write);
198
199 crc = (char *)&c_tx->pkt.write_pkt.crc;
200 data = siw_try_1seg(c_tx, crc);
201 break;
202
203 case SIW_OP_READ_RESPONSE:
204 memcpy(&c_tx->pkt.ctrl,
205 &iwarp_pktinfo[RDMAP_RDMA_READ_RESP].ctrl,
206 sizeof(struct iwarp_ctrl));
207
208 /* NBO */
209 c_tx->pkt.rresp.sink_stag = cpu_to_be32(wqe->sqe.rkey);
210 c_tx->pkt.rresp.sink_to = cpu_to_be64(wqe->sqe.raddr);
211
212 c_tx->ctrl_len = sizeof(struct iwarp_rdma_rresp);
213
214 crc = (char *)&c_tx->pkt.write_pkt.crc;
215 data = siw_try_1seg(c_tx, crc);
216 break;
217
218 default:
219 siw_dbg_qp(tx_qp(c_tx), "stale wqe type %d\n", tx_type(wqe));
220 return -EOPNOTSUPP;
221 }
222 if (unlikely(data < 0))
223 return data;
224
225 c_tx->ctrl_sent = 0;
226
227 if (data <= MAX_HDR_INLINE) {
228 if (data) {
229 wqe->processed = data;
230
231 c_tx->pkt.ctrl.mpa_len =
232 htons(c_tx->ctrl_len + data - MPA_HDR_SIZE);
233
234 /* Add pad, if needed */
235 data += -(int)data & 0x3;
236 /* advance CRC location after payload */
237 crc += data;
238 c_tx->ctrl_len += data;
239
240 if (!(c_tx->pkt.ctrl.ddp_rdmap_ctrl & DDP_FLAG_TAGGED))
241 c_tx->pkt.c_untagged.ddp_mo = 0;
242 else
243 c_tx->pkt.c_tagged.ddp_to =
244 cpu_to_be64(wqe->sqe.raddr);
245 }
246
247 *(u32 *)crc = 0;
248 /*
249 * Do complete CRC if enabled and short packet
250 */
251 if (c_tx->mpa_crc_hd &&
252 crypto_shash_digest(c_tx->mpa_crc_hd, (u8 *)&c_tx->pkt,
253 c_tx->ctrl_len, (u8 *)crc) != 0)
254 return -EINVAL;
255 c_tx->ctrl_len += MPA_CRC_SIZE;
256
257 return PKT_COMPLETE;
258 }
259 c_tx->ctrl_len += MPA_CRC_SIZE;
260 c_tx->sge_idx = 0;
261 c_tx->sge_off = 0;
262 c_tx->pbl_idx = 0;
263
264 /*
265 * Allow direct sending out of user buffer if WR is non signalled
266 * and payload is over threshold.
267 * Per RDMA verbs, the application should not change the send buffer
268 * until the work completed. In iWarp, work completion is only
269 * local delivery to TCP. TCP may reuse the buffer for
270 * retransmission. Changing unsent data also breaks the CRC,
271 * if applied.
272 */
273 if (c_tx->zcopy_tx && wqe->bytes >= SENDPAGE_THRESH &&
274 !(tx_flags(wqe) & SIW_WQE_SIGNALLED))
275 c_tx->use_sendpage = 1;
276 else
277 c_tx->use_sendpage = 0;
278
279 return PKT_FRAGMENTED;
280 }
281
282 /*
283 * Send out one complete control type FPDU, or header of FPDU carrying
284 * data. Used for fixed sized packets like Read.Requests or zero length
285 * SENDs, WRITEs, READ.Responses, or header only.
286 */
siw_tx_ctrl(struct siw_iwarp_tx * c_tx,struct socket * s,int flags)287 static int siw_tx_ctrl(struct siw_iwarp_tx *c_tx, struct socket *s,
288 int flags)
289 {
290 struct msghdr msg = { .msg_flags = flags };
291 struct kvec iov = { .iov_base =
292 (char *)&c_tx->pkt.ctrl + c_tx->ctrl_sent,
293 .iov_len = c_tx->ctrl_len - c_tx->ctrl_sent };
294
295 int rv = kernel_sendmsg(s, &msg, &iov, 1, iov.iov_len);
296
297 if (rv >= 0) {
298 c_tx->ctrl_sent += rv;
299
300 if (c_tx->ctrl_sent == c_tx->ctrl_len)
301 rv = 0;
302 else
303 rv = -EAGAIN;
304 }
305 return rv;
306 }
307
308 /*
309 * 0copy TCP transmit interface: Use MSG_SPLICE_PAGES.
310 *
311 * Using sendpage to push page by page appears to be less efficient
312 * than using sendmsg, even if data are copied.
313 *
314 * A general performance limitation might be the extra four bytes
315 * trailer checksum segment to be pushed after user data.
316 */
siw_tcp_sendpages(struct socket * s,struct page ** page,int offset,size_t size)317 static int siw_tcp_sendpages(struct socket *s, struct page **page, int offset,
318 size_t size)
319 {
320 struct bio_vec bvec;
321 struct msghdr msg = {
322 .msg_flags = (MSG_MORE | MSG_DONTWAIT | MSG_SPLICE_PAGES),
323 };
324 struct sock *sk = s->sk;
325 int i = 0, rv = 0, sent = 0;
326
327 while (size) {
328 size_t bytes = min_t(size_t, PAGE_SIZE - offset, size);
329
330 if (size + offset <= PAGE_SIZE)
331 msg.msg_flags &= ~MSG_MORE;
332
333 tcp_rate_check_app_limited(sk);
334 if (!sendpage_ok(page[i]))
335 msg.msg_flags &= ~MSG_SPLICE_PAGES;
336 bvec_set_page(&bvec, page[i], bytes, offset);
337 iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, &bvec, 1, bytes);
338
339 try_page_again:
340 lock_sock(sk);
341 rv = tcp_sendmsg_locked(sk, &msg, bytes);
342 release_sock(sk);
343
344 if (rv > 0) {
345 size -= rv;
346 sent += rv;
347 if (rv != bytes) {
348 bytes -= rv;
349 goto try_page_again;
350 }
351 offset = 0;
352 } else {
353 if (rv == -EAGAIN || rv == 0)
354 break;
355 return rv;
356 }
357 i++;
358 }
359 return sent;
360 }
361
362 /*
363 * siw_0copy_tx()
364 *
365 * Pushes list of pages to TCP socket. If pages from multiple
366 * SGE's, all referenced pages of each SGE are pushed in one
367 * shot.
368 */
siw_0copy_tx(struct socket * s,struct page ** page,struct siw_sge * sge,unsigned int offset,unsigned int size)369 static int siw_0copy_tx(struct socket *s, struct page **page,
370 struct siw_sge *sge, unsigned int offset,
371 unsigned int size)
372 {
373 int i = 0, sent = 0, rv;
374 int sge_bytes = min(sge->length - offset, size);
375
376 offset = (sge->laddr + offset) & ~PAGE_MASK;
377
378 while (sent != size) {
379 rv = siw_tcp_sendpages(s, &page[i], offset, sge_bytes);
380 if (rv >= 0) {
381 sent += rv;
382 if (size == sent || sge_bytes > rv)
383 break;
384
385 i += PAGE_ALIGN(sge_bytes + offset) >> PAGE_SHIFT;
386 sge++;
387 sge_bytes = min(sge->length, size - sent);
388 offset = sge->laddr & ~PAGE_MASK;
389 } else {
390 sent = rv;
391 break;
392 }
393 }
394 return sent;
395 }
396
397 #define MAX_TRAILER (MPA_CRC_SIZE + 4)
398
siw_unmap_pages(struct kvec * iov,unsigned long kmap_mask,int len)399 static void siw_unmap_pages(struct kvec *iov, unsigned long kmap_mask, int len)
400 {
401 int i;
402
403 /*
404 * Work backwards through the array to honor the kmap_local_page()
405 * ordering requirements.
406 */
407 for (i = (len-1); i >= 0; i--) {
408 if (kmap_mask & BIT(i)) {
409 unsigned long addr = (unsigned long)iov[i].iov_base;
410
411 kunmap_local((void *)(addr & PAGE_MASK));
412 }
413 }
414 }
415
416 /*
417 * siw_tx_hdt() tries to push a complete packet to TCP where all
418 * packet fragments are referenced by the elements of one iovec.
419 * For the data portion, each involved page must be referenced by
420 * one extra element. All sge's data can be non-aligned to page
421 * boundaries. Two more elements are referencing iWARP header
422 * and trailer:
423 * MAX_ARRAY = 64KB/PAGE_SIZE + 1 + (2 * (SIW_MAX_SGE - 1) + HDR + TRL
424 */
425 #define MAX_ARRAY ((0xffff / PAGE_SIZE) + 1 + (2 * (SIW_MAX_SGE - 1) + 2))
426
427 /*
428 * Write out iov referencing hdr, data and trailer of current FPDU.
429 * Update transmit state dependent on write return status
430 */
siw_tx_hdt(struct siw_iwarp_tx * c_tx,struct socket * s)431 static int siw_tx_hdt(struct siw_iwarp_tx *c_tx, struct socket *s)
432 {
433 struct siw_wqe *wqe = &c_tx->wqe_active;
434 struct siw_sge *sge = &wqe->sqe.sge[c_tx->sge_idx];
435 struct kvec iov[MAX_ARRAY];
436 struct page *page_array[MAX_ARRAY];
437 struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_EOR };
438
439 int seg = 0, do_crc = c_tx->do_crc, is_kva = 0, rv;
440 unsigned int data_len = c_tx->bytes_unsent, hdr_len = 0, trl_len = 0,
441 sge_off = c_tx->sge_off, sge_idx = c_tx->sge_idx,
442 pbl_idx = c_tx->pbl_idx;
443 unsigned long kmap_mask = 0L;
444
445 if (c_tx->state == SIW_SEND_HDR) {
446 if (c_tx->use_sendpage) {
447 rv = siw_tx_ctrl(c_tx, s, MSG_DONTWAIT | MSG_MORE);
448 if (rv)
449 goto done;
450
451 c_tx->state = SIW_SEND_DATA;
452 } else {
453 iov[0].iov_base =
454 (char *)&c_tx->pkt.ctrl + c_tx->ctrl_sent;
455 iov[0].iov_len = hdr_len =
456 c_tx->ctrl_len - c_tx->ctrl_sent;
457 seg = 1;
458 }
459 }
460
461 wqe->processed += data_len;
462
463 while (data_len) { /* walk the list of SGE's */
464 unsigned int sge_len = min(sge->length - sge_off, data_len);
465 unsigned int fp_off = (sge->laddr + sge_off) & ~PAGE_MASK;
466 struct siw_mem *mem;
467
468 if (!(tx_flags(wqe) & SIW_WQE_INLINE)) {
469 mem = wqe->mem[sge_idx];
470 is_kva = mem->mem_obj == NULL ? 1 : 0;
471 } else {
472 is_kva = 1;
473 }
474 if (is_kva && !c_tx->use_sendpage) {
475 /*
476 * tx from kernel virtual address: either inline data
477 * or memory region with assigned kernel buffer
478 */
479 iov[seg].iov_base =
480 ib_virt_dma_to_ptr(sge->laddr + sge_off);
481 iov[seg].iov_len = sge_len;
482
483 if (do_crc)
484 crypto_shash_update(c_tx->mpa_crc_hd,
485 iov[seg].iov_base,
486 sge_len);
487 sge_off += sge_len;
488 data_len -= sge_len;
489 seg++;
490 goto sge_done;
491 }
492
493 while (sge_len) {
494 size_t plen = min((int)PAGE_SIZE - fp_off, sge_len);
495 void *kaddr;
496
497 if (!is_kva) {
498 struct page *p;
499
500 p = siw_get_page(mem, sge, sge_off, &pbl_idx);
501 if (unlikely(!p)) {
502 siw_unmap_pages(iov, kmap_mask, seg);
503 wqe->processed -= c_tx->bytes_unsent;
504 rv = -EFAULT;
505 goto done_crc;
506 }
507 page_array[seg] = p;
508
509 if (!c_tx->use_sendpage) {
510 void *kaddr = kmap_local_page(p);
511
512 /* Remember for later kunmap() */
513 kmap_mask |= BIT(seg);
514 iov[seg].iov_base = kaddr + fp_off;
515 iov[seg].iov_len = plen;
516
517 if (do_crc)
518 crypto_shash_update(
519 c_tx->mpa_crc_hd,
520 iov[seg].iov_base,
521 plen);
522 } else if (do_crc) {
523 kaddr = kmap_local_page(p);
524 crypto_shash_update(c_tx->mpa_crc_hd,
525 kaddr + fp_off,
526 plen);
527 kunmap_local(kaddr);
528 }
529 } else {
530 /*
531 * Cast to an uintptr_t to preserve all 64 bits
532 * in sge->laddr.
533 */
534 u64 va = sge->laddr + sge_off;
535
536 page_array[seg] = ib_virt_dma_to_page(va);
537 if (do_crc)
538 crypto_shash_update(
539 c_tx->mpa_crc_hd,
540 ib_virt_dma_to_ptr(va),
541 plen);
542 }
543
544 sge_len -= plen;
545 sge_off += plen;
546 data_len -= plen;
547 fp_off = 0;
548
549 if (++seg >= (int)MAX_ARRAY) {
550 siw_dbg_qp(tx_qp(c_tx), "to many fragments\n");
551 siw_unmap_pages(iov, kmap_mask, seg-1);
552 wqe->processed -= c_tx->bytes_unsent;
553 rv = -EMSGSIZE;
554 goto done_crc;
555 }
556 }
557 sge_done:
558 /* Update SGE variables at end of SGE */
559 if (sge_off == sge->length &&
560 (data_len != 0 || wqe->processed < wqe->bytes)) {
561 sge_idx++;
562 sge++;
563 sge_off = 0;
564 }
565 }
566 /* trailer */
567 if (likely(c_tx->state != SIW_SEND_TRAILER)) {
568 iov[seg].iov_base = &c_tx->trailer.pad[4 - c_tx->pad];
569 iov[seg].iov_len = trl_len = MAX_TRAILER - (4 - c_tx->pad);
570 } else {
571 iov[seg].iov_base = &c_tx->trailer.pad[c_tx->ctrl_sent];
572 iov[seg].iov_len = trl_len = MAX_TRAILER - c_tx->ctrl_sent;
573 }
574
575 if (c_tx->pad) {
576 *(u32 *)c_tx->trailer.pad = 0;
577 if (do_crc)
578 crypto_shash_update(c_tx->mpa_crc_hd,
579 (u8 *)&c_tx->trailer.crc - c_tx->pad,
580 c_tx->pad);
581 }
582 if (!c_tx->mpa_crc_hd)
583 c_tx->trailer.crc = 0;
584 else if (do_crc)
585 crypto_shash_final(c_tx->mpa_crc_hd, (u8 *)&c_tx->trailer.crc);
586
587 data_len = c_tx->bytes_unsent;
588
589 if (c_tx->use_sendpage) {
590 rv = siw_0copy_tx(s, page_array, &wqe->sqe.sge[c_tx->sge_idx],
591 c_tx->sge_off, data_len);
592 if (rv == data_len) {
593 rv = kernel_sendmsg(s, &msg, &iov[seg], 1, trl_len);
594 if (rv > 0)
595 rv += data_len;
596 else
597 rv = data_len;
598 }
599 } else {
600 rv = kernel_sendmsg(s, &msg, iov, seg + 1,
601 hdr_len + data_len + trl_len);
602 siw_unmap_pages(iov, kmap_mask, seg);
603 }
604 if (rv < (int)hdr_len) {
605 /* Not even complete hdr pushed or negative rv */
606 wqe->processed -= data_len;
607 if (rv >= 0) {
608 c_tx->ctrl_sent += rv;
609 rv = -EAGAIN;
610 }
611 goto done_crc;
612 }
613 rv -= hdr_len;
614
615 if (rv >= (int)data_len) {
616 /* all user data pushed to TCP or no data to push */
617 if (data_len > 0 && wqe->processed < wqe->bytes) {
618 /* Save the current state for next tx */
619 c_tx->sge_idx = sge_idx;
620 c_tx->sge_off = sge_off;
621 c_tx->pbl_idx = pbl_idx;
622 }
623 rv -= data_len;
624
625 if (rv == trl_len) /* all pushed */
626 rv = 0;
627 else {
628 c_tx->state = SIW_SEND_TRAILER;
629 c_tx->ctrl_len = MAX_TRAILER;
630 c_tx->ctrl_sent = rv + 4 - c_tx->pad;
631 c_tx->bytes_unsent = 0;
632 rv = -EAGAIN;
633 }
634
635 } else if (data_len > 0) {
636 /* Maybe some user data pushed to TCP */
637 c_tx->state = SIW_SEND_DATA;
638 wqe->processed -= data_len - rv;
639
640 if (rv) {
641 /*
642 * Some bytes out. Recompute tx state based
643 * on old state and bytes pushed
644 */
645 unsigned int sge_unsent;
646
647 c_tx->bytes_unsent -= rv;
648 sge = &wqe->sqe.sge[c_tx->sge_idx];
649 sge_unsent = sge->length - c_tx->sge_off;
650
651 while (sge_unsent <= rv) {
652 rv -= sge_unsent;
653 c_tx->sge_idx++;
654 c_tx->sge_off = 0;
655 sge++;
656 sge_unsent = sge->length;
657 }
658 c_tx->sge_off += rv;
659 }
660 rv = -EAGAIN;
661 }
662 done_crc:
663 c_tx->do_crc = 0;
664 done:
665 return rv;
666 }
667
siw_update_tcpseg(struct siw_iwarp_tx * c_tx,struct socket * s)668 static void siw_update_tcpseg(struct siw_iwarp_tx *c_tx,
669 struct socket *s)
670 {
671 struct tcp_sock *tp = tcp_sk(s->sk);
672
673 if (tp->gso_segs) {
674 if (c_tx->gso_seg_limit == 0)
675 c_tx->tcp_seglen = tp->mss_cache * tp->gso_segs;
676 else
677 c_tx->tcp_seglen =
678 tp->mss_cache *
679 min_t(u16, c_tx->gso_seg_limit, tp->gso_segs);
680 } else {
681 c_tx->tcp_seglen = tp->mss_cache;
682 }
683 /* Loopback may give odd numbers */
684 c_tx->tcp_seglen &= 0xfffffff8;
685 }
686
687 /*
688 * siw_prepare_fpdu()
689 *
690 * Prepares transmit context to send out one FPDU if FPDU will contain
691 * user data and user data are not immediate data.
692 * Computes maximum FPDU length to fill up TCP MSS if possible.
693 *
694 * @qp: QP from which to transmit
695 * @wqe: Current WQE causing transmission
696 *
697 * TODO: Take into account real available sendspace on socket
698 * to avoid header misalignment due to send pausing within
699 * fpdu transmission
700 */
siw_prepare_fpdu(struct siw_qp * qp,struct siw_wqe * wqe)701 static void siw_prepare_fpdu(struct siw_qp *qp, struct siw_wqe *wqe)
702 {
703 struct siw_iwarp_tx *c_tx = &qp->tx_ctx;
704 int data_len;
705
706 c_tx->ctrl_len =
707 iwarp_pktinfo[__rdmap_get_opcode(&c_tx->pkt.ctrl)].hdr_len;
708 c_tx->ctrl_sent = 0;
709
710 /*
711 * Update target buffer offset if any
712 */
713 if (!(c_tx->pkt.ctrl.ddp_rdmap_ctrl & DDP_FLAG_TAGGED))
714 /* Untagged message */
715 c_tx->pkt.c_untagged.ddp_mo = cpu_to_be32(wqe->processed);
716 else /* Tagged message */
717 c_tx->pkt.c_tagged.ddp_to =
718 cpu_to_be64(wqe->sqe.raddr + wqe->processed);
719
720 data_len = wqe->bytes - wqe->processed;
721 if (data_len + c_tx->ctrl_len + MPA_CRC_SIZE > c_tx->tcp_seglen) {
722 /* Trim DDP payload to fit into current TCP segment */
723 data_len = c_tx->tcp_seglen - (c_tx->ctrl_len + MPA_CRC_SIZE);
724 c_tx->pkt.ctrl.ddp_rdmap_ctrl &= ~DDP_FLAG_LAST;
725 c_tx->pad = 0;
726 } else {
727 c_tx->pkt.ctrl.ddp_rdmap_ctrl |= DDP_FLAG_LAST;
728 c_tx->pad = -data_len & 0x3;
729 }
730 c_tx->bytes_unsent = data_len;
731
732 c_tx->pkt.ctrl.mpa_len =
733 htons(c_tx->ctrl_len + data_len - MPA_HDR_SIZE);
734
735 /*
736 * Init MPA CRC computation
737 */
738 if (c_tx->mpa_crc_hd) {
739 crypto_shash_init(c_tx->mpa_crc_hd);
740 crypto_shash_update(c_tx->mpa_crc_hd, (u8 *)&c_tx->pkt,
741 c_tx->ctrl_len);
742 c_tx->do_crc = 1;
743 }
744 }
745
746 /*
747 * siw_check_sgl_tx()
748 *
749 * Check permissions for a list of SGE's (SGL).
750 * A successful check will have all memory referenced
751 * for transmission resolved and assigned to the WQE.
752 *
753 * @pd: Protection Domain SGL should belong to
754 * @wqe: WQE to be checked
755 * @perms: requested access permissions
756 *
757 */
758
siw_check_sgl_tx(struct ib_pd * pd,struct siw_wqe * wqe,enum ib_access_flags perms)759 static int siw_check_sgl_tx(struct ib_pd *pd, struct siw_wqe *wqe,
760 enum ib_access_flags perms)
761 {
762 struct siw_sge *sge = &wqe->sqe.sge[0];
763 int i, len, num_sge = wqe->sqe.num_sge;
764
765 if (unlikely(num_sge > SIW_MAX_SGE))
766 return -EINVAL;
767
768 for (i = 0, len = 0; num_sge; num_sge--, i++, sge++) {
769 /*
770 * rdma verbs: do not check stag for a zero length sge
771 */
772 if (sge->length) {
773 int rv = siw_check_sge(pd, sge, &wqe->mem[i], perms, 0,
774 sge->length);
775
776 if (unlikely(rv != E_ACCESS_OK))
777 return rv;
778 }
779 len += sge->length;
780 }
781 return len;
782 }
783
784 /*
785 * siw_qp_sq_proc_tx()
786 *
787 * Process one WQE which needs transmission on the wire.
788 */
siw_qp_sq_proc_tx(struct siw_qp * qp,struct siw_wqe * wqe)789 static int siw_qp_sq_proc_tx(struct siw_qp *qp, struct siw_wqe *wqe)
790 {
791 struct siw_iwarp_tx *c_tx = &qp->tx_ctx;
792 struct socket *s = qp->attrs.sk;
793 int rv = 0, burst_len = qp->tx_ctx.burst;
794 enum rdmap_ecode ecode = RDMAP_ECODE_CATASTROPHIC_STREAM;
795
796 if (unlikely(wqe->wr_status == SIW_WR_IDLE))
797 return 0;
798
799 if (!burst_len)
800 burst_len = SQ_USER_MAXBURST;
801
802 if (wqe->wr_status == SIW_WR_QUEUED) {
803 if (!(wqe->sqe.flags & SIW_WQE_INLINE)) {
804 if (tx_type(wqe) == SIW_OP_READ_RESPONSE)
805 wqe->sqe.num_sge = 1;
806
807 if (tx_type(wqe) != SIW_OP_READ &&
808 tx_type(wqe) != SIW_OP_READ_LOCAL_INV) {
809 /*
810 * Reference memory to be tx'd w/o checking
811 * access for LOCAL_READ permission, since
812 * not defined in RDMA core.
813 */
814 rv = siw_check_sgl_tx(qp->pd, wqe, 0);
815 if (rv < 0) {
816 if (tx_type(wqe) ==
817 SIW_OP_READ_RESPONSE)
818 ecode = siw_rdmap_error(-rv);
819 rv = -EINVAL;
820 goto tx_error;
821 }
822 wqe->bytes = rv;
823 } else {
824 wqe->bytes = 0;
825 }
826 } else {
827 wqe->bytes = wqe->sqe.sge[0].length;
828 if (!rdma_is_kernel_res(&qp->base_qp.res)) {
829 if (wqe->bytes > SIW_MAX_INLINE) {
830 rv = -EINVAL;
831 goto tx_error;
832 }
833 wqe->sqe.sge[0].laddr =
834 (u64)(uintptr_t)&wqe->sqe.sge[1];
835 }
836 }
837 wqe->wr_status = SIW_WR_INPROGRESS;
838 wqe->processed = 0;
839
840 siw_update_tcpseg(c_tx, s);
841
842 rv = siw_qp_prepare_tx(c_tx);
843 if (rv == PKT_FRAGMENTED) {
844 c_tx->state = SIW_SEND_HDR;
845 siw_prepare_fpdu(qp, wqe);
846 } else if (rv == PKT_COMPLETE) {
847 c_tx->state = SIW_SEND_SHORT_FPDU;
848 } else {
849 goto tx_error;
850 }
851 }
852
853 next_segment:
854 siw_dbg_qp(qp, "wr type %d, state %d, data %u, sent %u, id %llx\n",
855 tx_type(wqe), wqe->wr_status, wqe->bytes, wqe->processed,
856 wqe->sqe.id);
857
858 if (--burst_len == 0) {
859 rv = -EINPROGRESS;
860 goto tx_done;
861 }
862 if (c_tx->state == SIW_SEND_SHORT_FPDU) {
863 enum siw_opcode tx_type = tx_type(wqe);
864 unsigned int msg_flags;
865
866 if (siw_sq_empty(qp) || !siw_tcp_nagle || burst_len == 1)
867 /*
868 * End current TCP segment, if SQ runs empty,
869 * or siw_tcp_nagle is not set, or we bail out
870 * soon due to no burst credit left.
871 */
872 msg_flags = MSG_DONTWAIT;
873 else
874 msg_flags = MSG_DONTWAIT | MSG_MORE;
875
876 rv = siw_tx_ctrl(c_tx, s, msg_flags);
877
878 if (!rv && tx_type != SIW_OP_READ &&
879 tx_type != SIW_OP_READ_LOCAL_INV)
880 wqe->processed = wqe->bytes;
881
882 goto tx_done;
883
884 } else {
885 rv = siw_tx_hdt(c_tx, s);
886 }
887 if (!rv) {
888 /*
889 * One segment sent. Processing completed if last
890 * segment, Do next segment otherwise.
891 */
892 if (unlikely(c_tx->tx_suspend)) {
893 /*
894 * Verbs, 6.4.: Try stopping sending after a full
895 * DDP segment if the connection goes down
896 * (== peer halfclose)
897 */
898 rv = -ECONNABORTED;
899 goto tx_done;
900 }
901 if (c_tx->pkt.ctrl.ddp_rdmap_ctrl & DDP_FLAG_LAST) {
902 siw_dbg_qp(qp, "WQE completed\n");
903 goto tx_done;
904 }
905 c_tx->state = SIW_SEND_HDR;
906
907 siw_update_tcpseg(c_tx, s);
908
909 siw_prepare_fpdu(qp, wqe);
910 goto next_segment;
911 }
912 tx_done:
913 qp->tx_ctx.burst = burst_len;
914 return rv;
915
916 tx_error:
917 if (ecode != RDMAP_ECODE_CATASTROPHIC_STREAM)
918 siw_init_terminate(qp, TERM_ERROR_LAYER_RDMAP,
919 RDMAP_ETYPE_REMOTE_PROTECTION, ecode, 1);
920 else
921 siw_init_terminate(qp, TERM_ERROR_LAYER_RDMAP,
922 RDMAP_ETYPE_CATASTROPHIC,
923 RDMAP_ECODE_UNSPECIFIED, 1);
924 return rv;
925 }
926
siw_fastreg_mr(struct ib_pd * pd,struct siw_sqe * sqe)927 static int siw_fastreg_mr(struct ib_pd *pd, struct siw_sqe *sqe)
928 {
929 struct ib_mr *base_mr = (struct ib_mr *)(uintptr_t)sqe->base_mr;
930 struct siw_device *sdev = to_siw_dev(pd->device);
931 struct siw_mem *mem;
932 int rv = 0;
933
934 siw_dbg_pd(pd, "STag 0x%08x\n", sqe->rkey);
935
936 if (unlikely(!base_mr)) {
937 pr_warn("siw: fastreg: STag 0x%08x unknown\n", sqe->rkey);
938 return -EINVAL;
939 }
940
941 if (unlikely(base_mr->rkey >> 8 != sqe->rkey >> 8)) {
942 pr_warn("siw: fastreg: STag 0x%08x: bad MR\n", sqe->rkey);
943 return -EINVAL;
944 }
945
946 mem = siw_mem_id2obj(sdev, sqe->rkey >> 8);
947 if (unlikely(!mem)) {
948 pr_warn("siw: fastreg: STag 0x%08x unknown\n", sqe->rkey);
949 return -EINVAL;
950 }
951
952 if (unlikely(mem->pd != pd)) {
953 pr_warn("siw: fastreg: PD mismatch\n");
954 rv = -EINVAL;
955 goto out;
956 }
957 if (unlikely(mem->stag_valid)) {
958 pr_warn("siw: fastreg: STag 0x%08x already valid\n", sqe->rkey);
959 rv = -EINVAL;
960 goto out;
961 }
962 /* Refresh STag since user may have changed key part */
963 mem->stag = sqe->rkey;
964 mem->perms = sqe->access;
965
966 siw_dbg_mem(mem, "STag 0x%08x now valid\n", sqe->rkey);
967 mem->va = base_mr->iova;
968 mem->stag_valid = 1;
969 out:
970 siw_mem_put(mem);
971 return rv;
972 }
973
siw_qp_sq_proc_local(struct siw_qp * qp,struct siw_wqe * wqe)974 static int siw_qp_sq_proc_local(struct siw_qp *qp, struct siw_wqe *wqe)
975 {
976 int rv;
977
978 switch (tx_type(wqe)) {
979 case SIW_OP_REG_MR:
980 rv = siw_fastreg_mr(qp->pd, &wqe->sqe);
981 break;
982
983 case SIW_OP_INVAL_STAG:
984 rv = siw_invalidate_stag(qp->pd, wqe->sqe.rkey);
985 break;
986
987 default:
988 rv = -EINVAL;
989 }
990 return rv;
991 }
992
993 /*
994 * siw_qp_sq_process()
995 *
996 * Core TX path routine for RDMAP/DDP/MPA using a TCP kernel socket.
997 * Sends RDMAP payload for the current SQ WR @wqe of @qp in one or more
998 * MPA FPDUs, each containing a DDP segment.
999 *
1000 * SQ processing may occur in user context as a result of posting
1001 * new WQE's or from siw_tx_thread context. Processing in
1002 * user context is limited to non-kernel verbs users.
1003 *
1004 * SQ processing may get paused anytime, possibly in the middle of a WR
1005 * or FPDU, if insufficient send space is available. SQ processing
1006 * gets resumed from siw_tx_thread, if send space becomes available again.
1007 *
1008 * Must be called with the QP state read-locked.
1009 *
1010 * Note:
1011 * An outbound RREQ can be satisfied by the corresponding RRESP
1012 * _before_ it gets assigned to the ORQ. This happens regularly
1013 * in RDMA READ via loopback case. Since both outbound RREQ and
1014 * inbound RRESP can be handled by the same CPU, locking the ORQ
1015 * is dead-lock prone and thus not an option. With that, the
1016 * RREQ gets assigned to the ORQ _before_ being sent - see
1017 * siw_activate_tx() - and pulled back in case of send failure.
1018 */
siw_qp_sq_process(struct siw_qp * qp)1019 int siw_qp_sq_process(struct siw_qp *qp)
1020 {
1021 struct siw_wqe *wqe = tx_wqe(qp);
1022 enum siw_opcode tx_type;
1023 unsigned long flags;
1024 int rv = 0;
1025
1026 siw_dbg_qp(qp, "enter for type %d\n", tx_type(wqe));
1027
1028 next_wqe:
1029 /*
1030 * Stop QP processing if SQ state changed
1031 */
1032 if (unlikely(qp->tx_ctx.tx_suspend)) {
1033 siw_dbg_qp(qp, "tx suspended\n");
1034 goto done;
1035 }
1036 tx_type = tx_type(wqe);
1037
1038 if (tx_type <= SIW_OP_READ_RESPONSE)
1039 rv = siw_qp_sq_proc_tx(qp, wqe);
1040 else
1041 rv = siw_qp_sq_proc_local(qp, wqe);
1042
1043 if (!rv) {
1044 /*
1045 * WQE processing done
1046 */
1047 switch (tx_type) {
1048 case SIW_OP_SEND:
1049 case SIW_OP_SEND_REMOTE_INV:
1050 case SIW_OP_WRITE:
1051 siw_wqe_put_mem(wqe, tx_type);
1052 fallthrough;
1053
1054 case SIW_OP_INVAL_STAG:
1055 case SIW_OP_REG_MR:
1056 if (tx_flags(wqe) & SIW_WQE_SIGNALLED)
1057 siw_sqe_complete(qp, &wqe->sqe, wqe->bytes,
1058 SIW_WC_SUCCESS);
1059 break;
1060
1061 case SIW_OP_READ:
1062 case SIW_OP_READ_LOCAL_INV:
1063 /*
1064 * already enqueued to ORQ queue
1065 */
1066 break;
1067
1068 case SIW_OP_READ_RESPONSE:
1069 siw_wqe_put_mem(wqe, tx_type);
1070 break;
1071
1072 default:
1073 WARN(1, "undefined WQE type %d\n", tx_type);
1074 rv = -EINVAL;
1075 goto done;
1076 }
1077
1078 spin_lock_irqsave(&qp->sq_lock, flags);
1079 wqe->wr_status = SIW_WR_IDLE;
1080 rv = siw_activate_tx(qp);
1081 spin_unlock_irqrestore(&qp->sq_lock, flags);
1082
1083 if (rv <= 0)
1084 goto done;
1085
1086 goto next_wqe;
1087
1088 } else if (rv == -EAGAIN) {
1089 siw_dbg_qp(qp, "sq paused: hd/tr %d of %d, data %d\n",
1090 qp->tx_ctx.ctrl_sent, qp->tx_ctx.ctrl_len,
1091 qp->tx_ctx.bytes_unsent);
1092 rv = 0;
1093 goto done;
1094 } else if (rv == -EINPROGRESS) {
1095 rv = siw_sq_start(qp);
1096 goto done;
1097 } else {
1098 /*
1099 * WQE processing failed.
1100 * Verbs 8.3.2:
1101 * o It turns any WQE into a signalled WQE.
1102 * o Local catastrophic error must be surfaced
1103 * o QP must be moved into Terminate state: done by code
1104 * doing socket state change processing
1105 *
1106 * o TODO: Termination message must be sent.
1107 * o TODO: Implement more precise work completion errors,
1108 * see enum ib_wc_status in ib_verbs.h
1109 */
1110 siw_dbg_qp(qp, "wqe type %d processing failed: %d\n",
1111 tx_type(wqe), rv);
1112
1113 spin_lock_irqsave(&qp->sq_lock, flags);
1114 /*
1115 * RREQ may have already been completed by inbound RRESP!
1116 */
1117 if ((tx_type == SIW_OP_READ ||
1118 tx_type == SIW_OP_READ_LOCAL_INV) && qp->attrs.orq_size) {
1119 /* Cleanup pending entry in ORQ */
1120 qp->orq_put--;
1121 qp->orq[qp->orq_put % qp->attrs.orq_size].flags = 0;
1122 }
1123 spin_unlock_irqrestore(&qp->sq_lock, flags);
1124 /*
1125 * immediately suspends further TX processing
1126 */
1127 if (!qp->tx_ctx.tx_suspend)
1128 siw_qp_cm_drop(qp, 0);
1129
1130 switch (tx_type) {
1131 case SIW_OP_SEND:
1132 case SIW_OP_SEND_REMOTE_INV:
1133 case SIW_OP_SEND_WITH_IMM:
1134 case SIW_OP_WRITE:
1135 case SIW_OP_READ:
1136 case SIW_OP_READ_LOCAL_INV:
1137 siw_wqe_put_mem(wqe, tx_type);
1138 fallthrough;
1139
1140 case SIW_OP_INVAL_STAG:
1141 case SIW_OP_REG_MR:
1142 siw_sqe_complete(qp, &wqe->sqe, wqe->bytes,
1143 SIW_WC_LOC_QP_OP_ERR);
1144
1145 siw_qp_event(qp, IB_EVENT_QP_FATAL);
1146
1147 break;
1148
1149 case SIW_OP_READ_RESPONSE:
1150 siw_dbg_qp(qp, "proc. read.response failed: %d\n", rv);
1151
1152 siw_qp_event(qp, IB_EVENT_QP_REQ_ERR);
1153
1154 siw_wqe_put_mem(wqe, SIW_OP_READ_RESPONSE);
1155
1156 break;
1157
1158 default:
1159 WARN(1, "undefined WQE type %d\n", tx_type);
1160 rv = -EINVAL;
1161 }
1162 wqe->wr_status = SIW_WR_IDLE;
1163 }
1164 done:
1165 return rv;
1166 }
1167
siw_sq_resume(struct siw_qp * qp)1168 static void siw_sq_resume(struct siw_qp *qp)
1169 {
1170 if (down_read_trylock(&qp->state_lock)) {
1171 if (likely(qp->attrs.state == SIW_QP_STATE_RTS &&
1172 !qp->tx_ctx.tx_suspend)) {
1173 int rv = siw_qp_sq_process(qp);
1174
1175 up_read(&qp->state_lock);
1176
1177 if (unlikely(rv < 0)) {
1178 siw_dbg_qp(qp, "SQ task failed: err %d\n", rv);
1179
1180 if (!qp->tx_ctx.tx_suspend)
1181 siw_qp_cm_drop(qp, 0);
1182 }
1183 } else {
1184 up_read(&qp->state_lock);
1185 }
1186 } else {
1187 siw_dbg_qp(qp, "Resume SQ while QP locked\n");
1188 }
1189 siw_qp_put(qp);
1190 }
1191
1192 struct tx_task_t {
1193 struct llist_head active;
1194 wait_queue_head_t waiting;
1195 };
1196
1197 static DEFINE_PER_CPU(struct tx_task_t, siw_tx_task_g);
1198
siw_create_tx_threads(void)1199 int siw_create_tx_threads(void)
1200 {
1201 int cpu, assigned = 0;
1202
1203 for_each_online_cpu(cpu) {
1204 struct tx_task_t *tx_task;
1205
1206 /* Skip HT cores */
1207 if (cpu % cpumask_weight(topology_sibling_cpumask(cpu)))
1208 continue;
1209
1210 tx_task = &per_cpu(siw_tx_task_g, cpu);
1211 init_llist_head(&tx_task->active);
1212 init_waitqueue_head(&tx_task->waiting);
1213
1214 siw_tx_thread[cpu] =
1215 kthread_run_on_cpu(siw_run_sq,
1216 (unsigned long *)(long)cpu,
1217 cpu, "siw_tx/%u");
1218 if (IS_ERR(siw_tx_thread[cpu])) {
1219 siw_tx_thread[cpu] = NULL;
1220 continue;
1221 }
1222 assigned++;
1223 }
1224 return assigned;
1225 }
1226
siw_stop_tx_threads(void)1227 void siw_stop_tx_threads(void)
1228 {
1229 int cpu;
1230
1231 for_each_possible_cpu(cpu) {
1232 if (siw_tx_thread[cpu]) {
1233 kthread_stop(siw_tx_thread[cpu]);
1234 wake_up(&per_cpu(siw_tx_task_g, cpu).waiting);
1235 siw_tx_thread[cpu] = NULL;
1236 }
1237 }
1238 }
1239
siw_run_sq(void * data)1240 int siw_run_sq(void *data)
1241 {
1242 const int nr_cpu = (unsigned int)(long)data;
1243 struct llist_node *active;
1244 struct siw_qp *qp;
1245 struct tx_task_t *tx_task = &per_cpu(siw_tx_task_g, nr_cpu);
1246
1247 while (1) {
1248 struct llist_node *fifo_list = NULL;
1249
1250 wait_event_interruptible(tx_task->waiting,
1251 !llist_empty(&tx_task->active) ||
1252 kthread_should_stop());
1253
1254 if (kthread_should_stop())
1255 break;
1256
1257 active = llist_del_all(&tx_task->active);
1258 /*
1259 * llist_del_all returns a list with newest entry first.
1260 * Re-order list for fairness among QP's.
1261 */
1262 fifo_list = llist_reverse_order(active);
1263 while (fifo_list) {
1264 qp = container_of(fifo_list, struct siw_qp, tx_list);
1265 fifo_list = llist_next(fifo_list);
1266 qp->tx_list.next = NULL;
1267
1268 siw_sq_resume(qp);
1269 }
1270 }
1271 active = llist_del_all(&tx_task->active);
1272 if (active) {
1273 llist_for_each_entry(qp, active, tx_list) {
1274 qp->tx_list.next = NULL;
1275 siw_sq_resume(qp);
1276 }
1277 }
1278 return 0;
1279 }
1280
siw_sq_start(struct siw_qp * qp)1281 int siw_sq_start(struct siw_qp *qp)
1282 {
1283 if (tx_wqe(qp)->wr_status == SIW_WR_IDLE)
1284 return 0;
1285
1286 if (unlikely(!cpu_online(qp->tx_cpu))) {
1287 siw_put_tx_cpu(qp->tx_cpu);
1288 qp->tx_cpu = siw_get_tx_cpu(qp->sdev);
1289 if (qp->tx_cpu < 0) {
1290 pr_warn("siw: no tx cpu available\n");
1291
1292 return -EIO;
1293 }
1294 }
1295 siw_qp_get(qp);
1296
1297 llist_add(&qp->tx_list, &per_cpu(siw_tx_task_g, qp->tx_cpu).active);
1298
1299 wake_up(&per_cpu(siw_tx_task_g, qp->tx_cpu).waiting);
1300
1301 return 0;
1302 }
1303